🧩 Tools & Plugins

Bundled tools become standalone plugins — Wait, notifications, search, A2A chat, scheduler, behavior adjustment, and the scheduler Tasks UI move into standard plugins with their own prompts and log styling.
Memory has one tool — The four memory operations now share a memory tool with an action argument. Legacy calls remain compatible; custom integrations should migrate to the unified tool.
Profiles control individual actions — Allow or deny operations across the 12 bundled multi-action tools, such as permitting file reads while blocking writes. The Agent Editor groups these controls in Easy and Advanced modes.
Tool calls are easier to recognize — Dedicated badges distinguish goal, document query, skills, vision, browser, office, and remote computer-use calls. All 47 bundled plugins also receive catalog thumbnails.

🏠 Welcome Screen & Search

Make the welcome screen your own — Dismiss sections, restore them together, and reorder blocks with mouse, touch, or keyboard. Layout choices persist in the browser, and plugins can add their own blocks.
Discover plugins from the start — A Plugin Hub button and channel-plugin cards bring discovery and settings to the welcome screen. System Resources and Quick Actions support half- and full-width layouts.
Find chats and messages — The sidebar offers Standard, Compact, and No search modes with keyboard navigation and Unicode-safe highlighting. Message search is also available through the Connector API.
Keep files in context — Each chat gets its own file-browser folder, history, and selection by default, with an option to share a browsing session across chats.

Responses update sooner — State pushes are coalesced at 25 milliseconds instead of 100, roughly halving the reported median interval between visible updates.
Tool calls recover more reliably — Output-limit truncation gets a dedicated retry warning, incomplete native arguments are rejected, repairable calls reach Context Doctor, and DSML envelopes go through normal dispatch.
Malformed output stays out of retries — Rejected or incomplete leaked calls are replaced with neutral history entries, while model prose and reasoning remain visible as thoughts during silent retries.
Tool arguments receive their included content — §§include(path) expands recursively at execution for Chat, native Responses, and parallel calls.
Chats resist rapid input and polling races — Duplicate sends are blocked, welcome-chat creation is serialized, and an in-flight pause is protected from stale snapshots. Message actions stay mounted during streaming.

Switch profiles while a chat runs — Agent profiles can be changed during execution, and the editor loads Advanced data only when needed. Prompt editing now shows clearer section names, customization status, save paths, and variable hints.
Chat names reflect real messages — The synthetic greeting no longer names a new chat, and subordinate agents receive the greeting's tool-call example. Matching specialists avoid redundant delegation to an identical profile.
Bring more history along — Migrate Agents can import Pi JSONL session versions 1–3 and preserve Claude chat titles and tool-result blocks.
Notifications and scheduled tasks recover better — Notifications can copy their title and message where appropriate, and scheduler errors clear after a successful continuation.
Setup and updates are more dependable — Provider discovery links to the Plugin Hub during onboarding, existing usr/.env settings remain authoritative during migration, and self-update has a longer health timeout and faster backups.

Browser DOM access is same-origin — The DOM helper rejects cross-origin windows. Host browsers must restart and reconnect to discard old initialization scripts.
Channel administration is explicit — Telegram and WhatsApp settings changes require configured administrators. Telegram configuration is now global-only; existing project- or profile-scoped settings pause the bot until reviewed and removed.
Remote input respects terminal permissions — input_remote now requires permission for remote terminal execution as well as its own permission.
Custom scheduler integrations may need adjustment — Initialize the scheduler store before reading snapshots. Development checkouts nested under /a0 also receive a one-time workspace ID change and new shadow-repository location.